CN101832471B - Signal identification and classification method - Google Patents

Abstract

The invention provides a signal identification and classification method. The method comprises the followings steps of: carrying out noise reduction on initial data containing higher noise by utilizing a wavelet transform method, decomposing signals into high-frequency information and low-frequency information in data analysis, carrying out noise cancelling on the signals by adopting a soft thresholding method and then carrying out signal reconstruction; carrying out further decomposition on the high-frequency part which is not detailedly classified by multiscale analysis while inheriting all the favorable time-frequency localization advantages of the wavelet transform; analyzing the signals within different frequency bands after multi-layered decomposition by utilizing the wavelet packet transform to extract out characteristic information reflecting a system state; transforming the characteristic vectors of input signals into a high-dimensional characteristic space through non-linear transform and then solving for an optimal linear classification plane in the high-dimensional characteristic space. The invention overcomes the defects of difficult determination of a network structure, low convergence rate, requirement on large quantities of data samples during training, and the like in neural network learning and enables the neural network learning to be with the characteristics of high precision and strong real time in the aspect of practical application of engineering.

Description

A kind of signal identification and classification method

Technical field

The present invention relates to the modern detecting field, specifically a kind of pipeline pressure detects and recognition technology.

Background technology

Since pipeline carry have cost low, save the energy, safe and supply with the advantages such as stable, pipeline transportation is worldwide developed rapidly, has become the indispensable ingredient of modern society.Yet, unaccelerated aging, geography and the reasons such as weather environmental evolution and artificial damage owing to long service wear, equipment, leakage failure happens occasionally, and has caused huge potential threat for people's life, property and living environment, also causes the serious wasting of resources simultaneously.Therefore, identifying timely and accurately Pipeline Leak has important practical significance.

Along with computing machine, signal are processed and the development of the technology such as pattern-recognition, received increasing concern based on the real-time leak detection technology of supervisory control and data acqui sition system (SCADA), and become gradually the main flow direction of detection technique development.These class methods mainly are the signals such as the pressure that gathers, flow to be carried out real-time analysis process, and detect and locate leakage point with this.Negative pressure wave method refers to that instantaneous pressure decreased appears in leak immediately when pipeline occurs to leak, and propagates with certain speed to the upstream and downstream of leakage point by pipeline and fluid media (medium) as the decompression wave source, and the decompression wave that produces during leakage just is called suction wave.The sensor that is arranged on the leakage point two ends is according to the variation of pressure signal and leak the mistiming that the suction wave that produces propagates into upstream and downstream, just can determine the leak position, and the method is sensitive and accurate, and principle is simple, and applicability is very strong.The automatic checkout system that China develops voluntarily adopts Negative-pressure Wave Principle that pipe leakage is detected more and locates, but bearing accuracy and automaticity all also have very large gap with external automatic checkout system.

Wavelet transformation is a kind of good signal analysis means that development in recent years is got up, it has good time frequency localization characteristic, can carry out multiscale analysis to signal by flexible and translation, can focus on any details of object, exactly realistic problem medium-high frequency is short signal duration for this, the long-term natural law of low frequency signal.A kind of new machine learning method-support vector machine (the Support vector machine that develops on the Statistical Learning Theory basis, SVM) practical problemss such as small sample, non-linear and higher-dimension pattern-recognition have been solved preferably, and overcome that network structure is difficult to determine in the network learning method, speed of convergence is slow, local minimum point, need the deficiency such as mass data sample when crossing study and owing study and training, has improved the Generalization Ability of learning method.Least square method supporting vector machine (Least Squares SupportVector Machines, LSSVM) be a kind of improvement of support vector machine, it is to change the inequality constrain in traditional support vector machine into equality constraint, to find the solution quadratic programming problem and be converted into the system of linear equations problem of finding the solution, improve speed and the convergence precision of Solve problems.The advantage of wavelet theory and LSSVM method is combined, can further improve the signal accuracy of identification.

Through existing technical literature retrieval is found, do not find and the same or similar bibliographical information of theme of the present invention.

Summary of the invention

The object of the present invention is to provide a kind of signal identification and classification method with high, the in real time strong characteristics of Engineering Oriented practical application precision.

The object of the present invention is achieved like this:

Step 1 wavelet transformation noise reduction at first utilizes the method for wavelet transformation that the raw data that contains higher noise is carried out noise reduction, is high and low frequency information with signal decomposition in data analysis, adopts the soft-threshold method that signal is carried out de-noising, then carries out signal reconstruction;

The step 2 WAVELET PACKET DECOMPOSITION in the good Time-Frequency Localization advantage that the succession wavelet transformation has, does not have the HFS of segmentation further to decompose to multiscale analysis;

The step 3 signal characteristic abstraction on the WAVELET PACKET DECOMPOSITION basis, utilizes the different frequency bands inner analysis signal of wavelet package transforms after multilayer is decomposed, and extracts the characteristic information of reflection system state.

Step 4 least square method supporting vector machine (Least Squares Support Vector Machines, LSSVM) identification, by nonlinear transformation with the input signal eigenvector transform to high-dimensional feature space, then ask for the optimum linearity classifying face at this high-dimensional feature space, this nonlinear transformation realizes by definition inner product function, as output, by training LSSVM, make the LSSVM network realize given input-output mappings relation normal condition and leak condition.

Principal feature of the present invention is embodied in:

(1) the employing wavelet transformation carries out the noise reduction of signal, is being mingled with signal and the noise of various frequencies in the Inhibitory signal, and this Method And Principle is simple, and it is convenient to realize, has effectively rejected signal noise.

(2) use wavelet packet to noise reduction after data decompose, extract proper vector, can in mass data, select can the characterization signal characteristics the minority vector, training data as least square method supporting vector machine, improved the work efficiency of system, laid a good foundation for training accurately the LSSVM network.

(3) can well to solve neural metwork training speed for the small sample training slow for least square method supporting vector machine, easily is absorbed in the shortcomings such as local extremum, and the solving-optimizing problem finally transfers to and find the solution system of linear equations, and computation process has obtained great simplification.

The present invention utilizes wavelet transformation to have pattern-recognition advantage under very strong capability of processing signals and least square method supporting vector machine small-sample learning and the multi-C vector space, propose a kind of pipeline based on wavelet transformation and least square method supporting vector machine and reveal recognition technology, overcome that network structure is difficult to determine in the neural network learning, speed of convergence is slow and has needed the deficiency such as mass data sample during training, made it have that Engineering Oriented practical application precision is high, real-time strong characteristics.

Description of drawings

Fig. 1 signal discriminator structural drawing;

Fig. 2 small echo one-dimensional signal noise reduction is figure as a result;

Three layers of WAVELET PACKET DECOMPOSITION schematic diagram of Fig. 3;

Fig. 4 support vector machine structural drawing.

Embodiment

For example the present invention is described in more detail below in conjunction with accompanying drawing:

In conjunction with Fig. 1, U1 is the wavelet transformation noise reduction, at first utilize the method for wavelet transformation that the raw data that contains higher noise is carried out noise reduction, in data analysis, be high and low frequency information with signal decomposition, in order to obtain preferably de-noising effect, adopt the soft-threshold method that signal is carried out de-noising, then carry out the effect that signal reconstruction can reach denoising; WAVELET PACKET DECOMPOSITION U2 is in the good Time-Frequency Localization advantage that the succession wavelet transformation has, there is not the HFS of segmentation further to decompose to multiscale analysis, thereby has better time-frequency characteristic, in theory, WAVELET PACKET DECOMPOSITION can ad infinitum go on, until only have the data of a point in the details of the bottom, but in actual applications, decide the number of plies of decomposition according to characteristics and the actual needs of signal; Signal characteristic abstraction U3 is the key point of signal identification success or not, and the feature extraction on the WAVELET PACKET DECOMPOSITION basis can capture the characteristic information that is hidden in the signal, and the signal characteristic information that extracts has sufficient susceptibility to different signals; Least square method supporting vector machine identification U4 by nonlinear transformation with the input signal eigenvector transform to high-dimensional feature space, then ask for the optimum linearity classifying face at this high-dimensional feature space, this nonlinear transformation realizes by defining the inner product function.

In conjunction with Fig. 2, in the signals such as pipeline pressure, flow, useful signal is usually expressed as more stably signal of low frequency signal or some, and noise then often shows as high-frequency signal, and the model of Noise one-dimensional signal can be expressed as following form:

s(k)＝f(k)+εe(k)，k＝0，1，...，n-1 (1)

Wherein, s (k) is signals and associated noises, and f (k) is useful signal, and e (k) is noise signal, and ε is the standard deviation of noise figure.

The basic thought of Wavelet Denoising Method is exactly that the wavelet coefficient on each yardstick has these characteristics of different manifestations according to noise and signal, with the Wavelet Component that is produced by noise on each yardstick, particularly the noise component on the prevailing yardstick of those noise components is removed, the wavelet coefficient that remains like this is exactly the wavelet coefficient of original signal basically, then utilize the wavelet reconstruction algorithm, reconstruct original signal.Adopt the db4 small echo to carry out the elimination of noise.Compare with wavelet basis function, decompose determining of the number of plies, larger for the influential effect of signal noise silencing, to decompose the number of plies the useful letter breath Lost of signal is lost, the very few noise reduction of the number of plies can descend.On the foundation of judging the signal decomposition number of plies, the pressure signal research according to for pipeline decomposes the 3rd layer for best, therefore, signal is carried out three layers of wavelet decomposition.On the choice problem of wavelet coefficient, the generic threshold value algorithm that adopts Donoho to propose.

The expression matrix form of wavelet decomposition algorithm recursion formula:

$\left\{\begin{array}{c}{C}_{j+1,k}=\underset{m}{\mathrm{\Σ}}{h}_0(m-2k)C_{j,m}\\ D_{j+1,k}=\underset{m}{\mathrm{\Σ}}{h}_1(m-2k)C_{j,m}\end{array}\right.---\left(2\right)$

Wherein, C _j, D _jRespectively by the discrete wavelet transformer coefficient of signal decomposition when becoming scaling function and wavelet function component of changing commanders, h ₀, h ₁Be respectively scaling function coefficient and wavelet function coefficient.

The reconstruct of signal is the inverse process of signal decomposition, the original coefficient of reconstruct, and corresponding reconstruction formula is then arranged:

$C_{j-1,k}=\underset{m}{\mathrm{\Σ}}{h}_0(k-2m)C_{j,m}+\underset{m}{\mathrm{\Σ}}{h}_1(k-2m)D_{j,m}---\left(3\right)$

In conjunction with Fig. 3, during pipe leakage, sudden change has occured in pressure signal and flow signal, WAVELET PACKET DECOMPOSITION is the popularization of multiscale analysis, signal is carried out meticulousr analysis, when having inherited the good Time-Frequency Localization advantage that wavelet transformation has, do not have the HFS of segmentation further to decompose to multiscale analysis, thereby have a better time-frequency characteristic.

For the standard Orthogonal Scaling Function function in the multiscale analysis

With wavelet function ψ, two scaling Equations are arranged:

H wherein _0kAnd h _1kBe respectively corresponding scaling function coefficient and wavelet function coefficient.

In wavelet multi-scale analysis, L ₂(R) spatial decomposition becomes the subspace { V that is comprised of scaling function _j, j ∈ Z and the subspace { W that is formed by wavelet function _j, j ∈ Z.The space that forms for wavelet packet is discussed, created symbol

With

By multiscale analysis as can be known It is generalized to wavelet packet, for ease of relatively, uses W _j ⁿExpression U _j ⁿ, so the wavelet packet spatial decomposition is:

$W_j^n=U_j^n=U_{j+1}^{2n}\⊕U_{j+1}^{2n+1}---\left(5\right)$

Because the n=0 correspondence wavelet decomposition, thus only consider n=1,2 ... and j=1,2 ..., recursion is gone down to such an extent that the general expression of WAVELET PACKET DECOMPOSITION is:

$W_j=U_{j+1}^2=U_{j+1}^3$

$W_j=U_{j+1}^4\⊕U_{j+1}^5\⊕U_{j+1}^6\⊕U_{j+1}^7---\left(6\right)$

……

$W_j=U_{j+k}^{2^k+1}\⊕U_{j+k}^{2^k+2}\⊕...\⊕U_{j+k}^{2^k-1}$

Spatial decomposition by wavelet packet can obtain, and the wavelet packet coefficient recursion formula is:

$d_k^{j+\mathrm{1,2}n}=\underset{l}{\mathrm{\Σ}}{h}_0(2l-k)d_n^{j,n}---\left(7\right)$

$d_k^{j+\mathrm{1,2}n+1}=\underset{l}{\mathrm{\Σ}}{h}_1(2l-k)d_n^{j,n}$

So the reconstruction formula of wavelet packet is:

$d_l^{j,n}=\underset{n}{\mathrm{\Σ}}[h_0(l-2k)d_k^{j+\mathrm{1,2}n}+h_1(l-2k)d_k^{j+\mathrm{1,2}n+1}]---\left(8\right)$

D wherein _k ^J+1,2nAnd d _k ^J+1,2n+1Be respectively signal at subspace U _J+1 ²ⁿAnd U _J+1 ²ⁿ⁺¹On wavelet packet coefficient.

Decide the number of plies of decomposition according to the characteristics of signal and actual needs, three layers of WAVELET PACKET DECOMPOSITION can meet the demands.Among Fig. 3, A represents low frequency, and D represents high frequency, the number of plies that numeral is decomposed, and therefore three layers of WAVELET PACKET DECOMPOSITION are closed and are:

S＝AAA ₃+DAA ₃+ADA ₃+DDA ₃+AAD ₃+DAD ₃+ADD ₃+DDD ₃

From the energy point of view analysis, when pipe leakage, sudden change has occured in pressure signal, and the space distribution of its output signal energy is compared with normal system output, and respective change can occur, and the change of namely exporting energy is comprising abundant signal intensity characteristic information.Therefore, if extract feature from the distribution of signal energy each subspace, namely utilize the different frequency bands inner analysis signal of wavelet package transforms after multilayer is decomposed, make that the form with significant energy variation shows in these unconspicuous signal frequency feature some subspaces on each yardstick, and compare with the normal output of system, extract the characteristic information of reflection system state.Basic step is:

(1) extract the 3rd layer in the signal characteristic of each frequency content, and establish S and represent original signal, extract the signal S of each frequency band range _Ij, the WAVELET PACKET DECOMPOSITION coefficient is D _Ij, i=0 wherein, 1,2,3, j=0,1 ..., 2 ³-1;

(2) ask the gross energy of each band signal, establish signal S _IjCorresponding energy is E _Ij, then have:

$E_{3j}={\∫\left|S_{3j}\left(t\right)\right|}^2\mathrm{dt}=\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{\left|x_{\mathrm{jk}}\right|}^2---\left(7\right)$

In the formula, x _JkExpression reconstruction signal S _IjThe discrete point amplitude, j=0,1 ..., 7, k=0,1 ..., n, n are the pressure signal sampling number.

(3) structural attitude vector.When pipeline occurs to leak, can larger impact be arranged to the energy of signal in each frequency band, therefore, proper vector T of structure represents the pressure characteristic signal take energy as element, and proper vector is carried out normalized:

$T=\frac{[E_{30},E_{31},...E_{37}]}{{\left(\underset{j=1}{\overset{7}{\mathrm{\Σ}}}{\left(E_{3j}\right)}^2\right)}^{\frac{1}{\text{2}}}}---\left(8\right)$

Wherein: E _IjBe signal S _IjCorresponding energy.

In conjunction with Fig. 4, least square method supporting vector machine is found the solution optimal classification face problem in the linear separability situation and is developed, by certain nonlinear transformation of selecting in advance input vector is mapped to a high-dimensional feature space, structure optimal classification lineoid in this feature space, and this nonlinear transformation is to be that kernel function realizes by defining suitable inner product function.Among the figure,

Be kernel function, { x _i, y _i, i=1 ..., n, x _i∈ R ^N, y _i∈ R.

The LSSVM classification function is similar to a neural network in form, output is the linear combination of intermediate node, each intermediate node is corresponding to a support vector, and adopt the least square linear system as loss function, the QUADRATIC PROGRAMMING METHOD FOR that replaces traditional support vector machine, be converted into the system of linear equations problem of finding the solution separating quadratic programming problem, improve speed and the convergence precision of Solve problems.

Pipeline pressure proper vector sample set is n sample { x _i, y _i, i=1 wherein ..., n, x _i∈ R ^N, { 1,1} is category label to y ∈.For pipeline pressure signal identification, adopt the forms of the single output of many inputs, as input, as output, by training LSSVM, make network realize that given input-output mappings concerns normal condition and leak condition the pipeline pressure signal characteristic vector.In higher dimensional space, the error-free classification lineoid that separates of normal condition and leak condition proper vector sample is satisfied:

ω∈R ^N，b∈R (9)

Wherein

Be nonlinear function, input is mapped to feature space, ω, b represent respectively weight coefficient and biasing.

The constant term b of LSSVM classification lineoid and Lagrange multiplier vector α are solved by system of linear equations, and the optimal classification function that obtains is:

$y\left(x\right)=\mathrm{sgn}\{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}K(x_i\·x)+b\}---\left(10\right)$

In the formula, sgn{} is sign function.

The advantage of least square method supporting vector machine method is to there is no need to know the concrete form of mapping, and only need define the inner product operation K (x in the higher dimensional space _i, x) (adopt radial basis kernel function K (x _i, x)=exp (|| x _i-x|| ²/ 2 σ ²), σ wherein ²Be cuclear density) get final product, even conversion rear space dimension increases much like this, the complexity of calculating does not have too large variation yet.

Claims (4)

1. signal identification and classification method is characterized in that:

Step 1: the pipeline pressure of supervisory control and data acqui sition system collection, flow signal at first carry out the wavelet transformation noise reduction, utilize the method for wavelet transformation that the raw data that contains higher noise is carried out noise reduction, in data analysis, be high and low frequency information with signal decomposition, adopt the soft-threshold method that signal is carried out de-noising, then carry out signal reconstruction;

Step 2: WAVELET PACKET DECOMPOSITION in the good Time-Frequency Localization advantage that the succession wavelet transformation has, does not have the HFS of segmentation further to decompose to multiscale analysis;

Step 3: signal characteristic abstraction, on the WAVELET PACKET DECOMPOSITION basis, utilize the different frequency bands inner analysis signal of wavelet package transforms after multilayer is decomposed, extract the characteristic information of reflection system state;

Step 4: least square method supporting vector machine identification, by nonlinear transformation with the input signal eigenvector transform to high-dimensional feature space, utilize the structure risk minimum principle, ask for the optimum linearity classifying face at this high-dimensional feature space, this nonlinear transformation realizes by definition inner product function.

2. a kind of signal identification and classification method according to claim 1, it is characterized in that: described WAVELET PACKET DECOMPOSITION is that signal is carried out three layers of wavelet decomposition, so the reconstruction formula of small echo is:

$C_{j-1,k}=\underset{m}{\mathrm{\Σ}}{h}_0(k-2m)C_{j,m}+\underset{m}{\mathrm{\Σ}}{h}_1(k-2m)D_{j,m}$

Wherein: C _j, D _jRespectively by the discrete wavelet transformer coefficient of signal decomposition when becoming scaling function and wavelet function component of changing commanders, h ₀, h ₁Be respectively scaling function coefficient and wavelet function coefficient.

3. a kind of signal identification and classification method according to claim 2, it is characterized in that: the basic step of described signal characteristic abstraction is:

(1) extract the 3rd layer in the signal characteristic of each frequency content, and establish S and represent original signal, extract the signal S of each frequency band range _Ij, the WAVELET PACKET DECOMPOSITION coefficient is D _Ij, i=0 wherein, 1,2,3, j=0,1 ..., 2 ³-1;

(2) ask the gross energy of each band signal, establish signal S _IjCorresponding energy is E _Ij, then have:

$E_{3j}={\∫\left|S_{3j}\left(t\right)\right|}^2\mathrm{dt}=\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{\left|x_{\mathrm{jl}}\right|}^2$

In the formula, x _JlExpression reconstruction signal S _3jThe discrete point amplitude, j=0,1 ..., 7, l=0,1 ..., n, n are the pressure signal sampling number;

(3) structural attitude vector, proper vector T of structure represents the pressure characteristic signal take energy as element, and proper vector is carried out normalized:

$T=\frac{[E_{30},E_{31},\·\·\·,E_{37}]}{{\left(\underset{j=1}{\overset{7}{\mathrm{\Σ}}}{\left(e_{3j}\right)}^2\right)}^{\frac{1}{2}}}$

Wherein: E _3jBe signal S _3jCorresponding energy.

4. a kind of signal identification and classification method according to claim 3 is characterized in that: the basic step of described least square method supporting vector machine identification is:

Least square method supporting vector machine identification, by nonlinear transformation with the input signal eigenvector transform to high-dimensional feature space, then ask for the optimum linearity classifying face at this high-dimensional feature space, this nonlinear transformation realizes by defining the inner product function; Pipeline pressure proper vector sample set is n sample { x _i, y _i, i=1 wherein ..., n, x _i∈ R ^N, { 1,1} is category label to y ∈; As input, as output, by training LSSVM, make the LSSVM network realize given input-output mappings relation normal condition and leak condition signal characteristic vector; In higher dimensional space, the error-free classification lineoid that separates of normal condition and leak condition proper vector sample is satisfied:

ω∈R ^N，b∈R

Wherein

Be nonlinear function, input is mapped to feature space, ω, b represent respectively weight coefficient and biasing;

The constant term b of LSSVM classification lineoid and Lagrange multiplier vector α are solved by system of linear equations, and the optimal classification function that obtains is:

$y\left(x\right)=\mathrm{sgn}\{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}K(x_i\·x)+b\}$

In the formula, sgn{} is sign function, adopts radial basis kernel function K (x _i, x)=exp (|| x _i-x|| ²/ 2 σ ²), σ wherein ²Be cuclear density.

Images